PASAformer: Cerebrovascular Disease Classification With Medical Prior-Guided Adapter and Pathology-Aware Sparse...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology and vascular medicine

Background:

• Cerebrovascular diseases (CVDs) pose significant public health challenges, necessitating accurate classification for effective treatment.
• Current computer-aided diagnosis (CAD) methods for CVDs struggle with limited representation, feature redundancy, and poor interpretability.
• Digital Subtraction Angiography (DSA) is a key imaging modality for diagnosing CVDs, but automated analysis remains challenging.

Purpose of the Study:

• To develop and evaluate PASAformer, a novel Swin-Transformer-based framework for automated classification of cerebrovascular diseases using DSA.
• To introduce a Pathology-Aware Sparse Attention (PASA) module to enhance focus on relevant pathological regions and improve computational efficiency.
• To establish CDSA-NEO, the first large-scale benchmark dataset for cerebrovascular disease classification from DSA images.

Main Methods:

• PASAformer utilizes a Swin-Transformer backbone integrated with a Pathology-Aware Sparse Attention (PASA) module, replacing standard self-attention for improved efficiency.
• The MiAMix data augmentation technique is employed to increase the diversity of the training samples.
• A CombinedAdapter encoder incorporates anatomical priors from the Medical Segment Anything Model (MED-SAM) to boost performance under limited supervision.

Main Results:

• PASAformer demonstrated competitive precision and balanced accuracy on the CDSA-NEO dataset and public vascular datasets compared to state-of-the-art models.
• The PASA module effectively emphasizes lesion-related regions and suppresses background noise, leading to more focused visual explanations.
• The framework showed robustness in realistic temporal workflows when evaluated on an external cohort of angiographic runs.

Conclusions:

• PASAformer offers a promising solution for automated cerebrovascular disease classification on DSA, improving upon existing CAD methods.
• The CDSA-NEO dataset serves as a valuable resource for advancing research and development in automated CVD analysis.
• The proposed framework has the potential to support clinical decision-making and improve patient outcomes through timely and accurate diagnosis.